TeamAI CLI: Tencent's Layer Above Your Agents

Tencent open-sourced TeamAI CLI at the end of April 2026, and the fastest way to misread it is to look at the name and assume Tencent shipped another coding agent. It didn't. TeamAI CLI doesn't write code, doesn't hold a conversation, and doesn't own a model. It sits one layer above Claude Code, Codex, Cursor and eight others, and its only job is making sure every person on a team is running the same skills, the same rules and the same MCP servers as everyone else.

That distinction matters more than it sounds. A team of eight people using Claude Code is really eight private setups that happen to share a repo. Someone writes a genuinely good review skill on a Tuesday, it lives in their ~/.claude folder, and nobody else ever sees it.

One git repo, eleven agents, one set of habits

The mechanism is deliberately boring: a git repo. An admin runs teamai init https://github.com/yourorg/yourrepo against a repo everyone has write access to, and that repo becomes the place where skills, rules, docs, agent definitions, hooks, MCP config and shared environment switches live. Members run the same init and pick a scope. Project scope drops everything under the current project directory; --scope user installs under ~/ instead, which is what you want if you hop between repos all day.

From then on the sync is automatic. TeamAI CLI registers a SessionStart hook, so when a teammate opens an agent session, teamai pull runs before they type anything and whatever the admins have published is already in place.

What actually lands on a teammate's machine

Each resource type has a fixed home in the team repo, which is the part that makes this reviewable rather than magic:

• Skills at skills/<name>/SKILL.md, rules at rules/*.md, and agents as agents/<name>.yaml
• culture.md, which holds the team's mission and working principles and gets injected into each agent's CLAUDE.md or AGENTS.md so every session inherits it
• docs/ for foundational project documentation, deliberately not all loaded by default
• hooks/hooks.yaml, mcp/mcp.yaml, and an env/ directory for shared team-level switches, with a blunt warning in the docs not to put secrets there
• teamai.yaml for packages, which today means npm packages and Claude Code plugins and nothing else

The culture.md idea is the one I'd have expected a Tencent internal tool to skip, and it's the one that will either be quietly useful or quietly ignored depending on whether your team writes down how it works. Models sit in that same resource table with a blank cell and a note that they are not implemented for every provider yet.

That warning on env/ is worth sitting with, because it draws the boundary of what any git-based distribution tool can do. Rules and skills travel; credentials, tokens and the half-configured local setup they plug into do not, which is exactly the part where onboarding a new teammate stalls for two days. What TeamAI CLI hands a new hire is a correctly populated set of instructions pointing at an environment they still have to build themselves. Zentor comes at that half from the other side: a hosted cloud AI computer that already exists and is already wired up, so the shared thing is the working machine rather than the instructions for making one.

The push, review, pull loop is the whole product

Publishing isn't a file copy. A member runs teamai push, which creates a branch and opens a merge request; a reviewer approves and merges; the next teamai pull on everyone else's machine picks it up. Skills go through code review, in other words, because they are code review artifacts now.

That's a real opinion baked into a tool, and it cuts both ways. It means nobody quietly ships a rule that tells every agent on the team to skip tests. It also means the fastest path from "I found a better prompt" to "everyone has it" runs through whatever your MR queue looks like on a bad week.

Git provider support is wider than the usual GitHub-only assumption: GitHub, GitLab, GitCode, CNB, TGit, or a private git service. Given Tencent's own internal use since March 2026, TGit showing up in that list is not a surprise.

Eleven agents, but the support matrix isn't flat

The README's compatibility table is the single most useful thing in the repo, and it's worth reading carefully rather than counting the logos. Eleven agents appear: Claude Code, Codex, Cursor, CodeBuddy, WorkBuddy, OpenCode, OpenClaw, Hermes, DeepSeek Harness, Qoder and ZCode.

Four of them get every box ticked across all three product layers. The rest have holes. WorkBuddy takes everything except agent definitions. OpenCode supports the whole execution and context surface but none of the Team Improvement features. OpenClaw takes skills, rules, docs and env, but not agents, hooks or MCP. Hermes is thinner still: skills, docs and env only. DeepSeek Harness is the narrowest supported agent on the list, receiving skills and docs and nothing else from the execution layer, which lines up with how plugins and skills work differently in that harness.

So "TeamAI CLI supports your agent" and "TeamAI CLI can distribute the thing you care about to your agent" are two different claims. Check the row, not the name.

Read the table long enough and the shape of the ceiling shows up: what a team can standardise is capped by whichever agent its people happen to have installed, and no amount of work on the distribution layer moves that. Capability that shouldn't depend on someone's local install has to live somewhere that isn't a local install, which is the argument for keeping a Zentor cloud AI computer in the mix alongside the laptops rather than trying to make eleven agents agree.

Underneath the CLI, Tencent describes three layers. Team Execution is shipping today and covers init, pull, push and all the resource types above. Team Context and Team Improvement are both marked beta, and cover recall, learnings, a codebase graph, a team wiki, session digests and a usage dashboard. Beta here appears to mean genuinely in progress rather than marketing caution.

Roles, tags and sources decide who gets which skills

Pushing every skill to every person stops working somewhere around the second team that adopts this. Three commands handle the fan-out. teamai roles maps roles to namespaces so a backend engineer doesn't sync the design team's skills. teamai tags lets people subscribe to specific tags. teamai source subscribes your repo to other repos — another team's public repo, or a shared internal one — and those skills sync on pull alongside your own.

That last one is the quiet architectural bet. It turns a single team repo into a small package ecosystem, and it's also the feature most likely to produce a mess if nobody owns the namespace.

The template library is currently one template

The README points newcomers at the teamai-hub org for a starting repo pre-loaded with skills, rules and review agents. Open it and there are two repositories and exactly one template: template-backend, described as skills, common rules and review agents for a backend engineer. The README says more roles are coming, naming frontend, QA, PM, data and design, and invites contributions.

That's not a criticism so much as a calibration. If you were planning to adopt TeamAI CLI by cloning a ready-made setup for your role, there's a good chance your role isn't there yet and you'll be writing the skills yourself. The 84KB usage guide in docs/usage-guide.md runs to about 1,570 lines and covers team creation through day-to-day use, so at least the writing-it-yourself path is documented.

Licence, version, and how new this really is

Check the LICENSE file rather than trusting an API. GitHub's licence detector reports NOASSERTION for this repo, which normally makes me suspicious, but here it's a false alarm caused by formatting: Tencent prepends its own paragraph before the licence text, and the file states plainly that teamai-cli is licensed under MIT with no additional restrictions beyond it. A commit five months ago is titled "update LICENSE to Tencent open source MIT license". It's MIT.

As of 10 September 2026 the repo carries 3,097 stars, 195 forks, 703 commits and 39 contributors, with 42 releases and the latest tagged v0.23.1 less than a day before this was written. Created 27 April 2026, so about four and a half months old. Twenty branches, 82 tags, 18 open issues. This is an actively maintained project with real outside contributors, not a code drop.

Version numbers still start with a zero, though, and the two beta layers are the half of the product that would justify adopting it over a shared dotfiles repo.

Who should install TeamAI CLI today

If your team is two people, skip it; a shared repo and a copy-paste habit will beat the merge-request overhead. Somewhere past five or six people running agents daily, the arithmetic flips, especially if half of them are on Claude Code and the other half on Cursor or Codex and nobody can remember whose rules are current. The plugin marketplace approach solves a related distribution problem for one agent; TeamAI CLI is aiming at the same problem across all of them at once.

One thing to be honest about: everything TeamAI CLI does terminates in files on someone's laptop, under a project directory or under ~/. New machine, new install. Machine asleep, no session, no pull. That's the correct design for a tool whose whole premise is local agents, and it's also exactly where a hosted environment earns its keep, because Zentor gives the team a cloud AI computer that isn't tied to whichever laptop happens to be open, and it sits beside your local setup rather than asking you to abandon it.

Install it with npm install -g teamai-cli. If you're evaluating, the honest test is whether your team already has three or four skills good enough that you'd be annoyed if only one person had them. If the answer is no, the distribution problem isn't your bottleneck yet, and a good place to start is a shortlist of skills worth sharing.

FAQ

Is TeamAI CLI free and open source?

Yes. The LICENSE file states MIT, and the repo is public under the Tencent org on GitHub. The npm package teamai-cli installs globally with no account or key. GitHub's automatic licence detection reports NOASSERTION because Tencent adds a custom preamble above the licence text, which is a formatting artifact and not a restriction.

Does TeamAI CLI replace Claude Code or Cursor?

No, and it can't. It has no model, no chat interface and no editing loop of its own. It distributes skills, rules, docs, agents, hooks and MCP config into the agents you already run, so you keep whichever one you use and TeamAI CLI changes what that agent knows about how your team works.

What is a teamai-hub template?

A GitHub template repo pre-loaded with a role's skills, rules and review agents. You click "Use this template" to create your own copy, then run teamai init against it instead of starting from an empty repo. As of September 2026 the teamai-hub org publishes one, template-backend, with frontend, QA, PM, data and design listed as planned.

Which coding agents does TeamAI CLI support?

Eleven are listed: Claude Code, Codex, Cursor, CodeBuddy, WorkBuddy, OpenCode, OpenClaw, Hermes, DeepSeek Harness, Qoder and ZCode. Support isn't uniform. Claude Code, Codex, Cursor, CodeBuddy and Qoder receive every resource type; Hermes and DeepSeek Harness receive only a subset, so check the README's compatibility table for the specific resource you plan to distribute.

Do team members need write access to the team repo?

Admins grant write access so members can teamai push, but a push opens a merge request rather than committing to the default branch, so review still gates what reaches everyone else. Read-only consumption works too if you only ever pull.